Abstract Fien de Mulder
Although a well-developed area for English (where high levels of accuracy are achieved), there hasn't been any work we are aware of on named-entity recognition (NER) for Dutch (with the notable exception of Buchholz and Van den Bosch, 2000).
In this paper,
(i) we describe an NER system for persons, companies, and locations,
based on gazetteers and simple "sure-fire" rules. Precision of this
baseline system on unseen parts of our corpus (a part from the Flemish
journal FET) is between 90 and 99% depending on the type of NE, with
low recall levels between 22 and 55%.
(ii) we show how this baseline system can be used to provide a seed
for the extraction of more sophisticated rules (using Ripper, a rule
induction algorithm), improving recall.
(iii) we describe an NP chunker consisting of a hand-crafted
regular expression part and a (memory-based) machine learning part.
(iv) we show the interaction between both components: the chunk
information can be used to improve NER, and the NER information can be
used to improve chunking.
Last modified $Date: 2001/10/04 13:39:44 $ by Parlevink Webmaster